编译｜未玖

Nature,  13 November 2025, Volume 647, Issue 8089 

《自然》，2025年11月13日，第647卷，8089期 

物理学Physics   

Millisecond lifetimes and coherence times in 2D transmon qubits 

二维透射子量子比特的毫秒寿命和相干时间    

▲ 作者：Matthew P. Bland, Faranak Bahrami, Jeronimo G. C. Martinez, Paal H. Prestegaard, Basil M. Smitham, Atharv Joshi, et al.

▲链接：

https://www.nature.com/articles/s41586-025-09687-4 

▲摘要：材料改进是减少超导量子比特损耗和退相干的有效途径，因为这种改进可较易转化为大规模处理器。最近的工作通过使用钽作为基底层和蓝宝石作为基底来改善透射子相干性。这些器件的损耗主要由两级系统控制，表面电介质和体电介质的贡献相当，这表明要实现技术水平的实质性改进，必须同时解决这两个问题。

研究组展示了用高电阻率硅代替基底显著降低了体基底损耗，在45个量子比特上实现了时间平均品质因子（Q^avg）高达9.7×10⁶的2D透射子。对于最佳量子比特，研究组实现了1.5×107的Q^avg，最大Q值达到2.5×10⁷，对应寿命（T₁）高达1.68 ms。这种低损耗也使研究组能够观察到与约瑟夫森结相关的退相干效应，并使用改进的低污染结沉积来实现超过T1的哈恩回波相干时间（T_2E）。 

研究组在不修改量子比特架构的情况下实现了这些材料的改进，从而实现轻松集成标准量子控制门。他们展示了具有99.994%保真度的单量子比特门。硅上钽平台包括一个可在晶圆规模上制造的简单材料堆叠结构，因此可便捷转化为大规模量子处理器。

▲ Abstract： Materials improvement is a powerful approach to reducing loss and decoherence in superconducting qubits, because such improvements can be readily translated to large-scale processors. Recent work improved transmon coherence by using tantalum as a base layer and sapphire as a substrate. The losses in these devices are dominated by two-level systems with comparable contributions from both the surface and bulk dielectrics, indicating that both must be tackled to achieve substantial improvements in the state of the art. Here we show that replacing the substrate with high-resistivity silicon markedly decreases the bulk substrate loss, enabling 2D transmons with time-averaged quality factors (Q^avg) of 9.7×10⁶ across 45qubits. For our best qubit, we achieve a Qavg of 1.5×107, reaching a maximum Q of 2.5×10⁷, corresponding to a lifetime (T₁) up to 1.68 ms. This low loss also allows us to observe decoherence effects related to the Josephson junction, and we use an improved, low-contamination junction deposition to achieve Hahn echo coherence times (T_2E) exceeding T1. We achieve these materials improvements without any modifications to the qubit architecture, allowing us to readily incorporate standard quantum control gates. We demonstrate single-qubit gates with 99.994% fidelity. The tantalum-on-silicon platform comprises a simple material stack that can potentially be fabricated at the wafer scale and therefore can be readily translated to large-scale quantum processors.   

人工智能Artificial Intelligence   

Aligning machine and human visual representations across abstraction levels 

跨抽象层级对齐机器和人类的视觉表征    

▲ 作者：Lukas Muttenthaler, Klaus Greff, Frieda Born, Bernhard Spitzer, Simon Kornblith, Michael C. Mozer, et al.

▲ 链接： https://www.nature.com/articles/s41586-025-09631-6 

▲摘要：深度神经网络已在广泛应用中取得了成功，包括作为人类行为模型和视觉任务中的神经表征。然而，神经网络训练和人类学习存在根本差异，神经网络通常不能像人类那样稳健泛化，这就引发了人们对其底层表征相似性的质疑。人们需要确定，现代学习系统要表现出更符合人类的行为，还欠缺什么？

研究组强调了视觉模型和人类之间的一个关键差异：尽管人类概念知识是从精细到粗粒度来层级化组织的，但模型表征并不能准确捕获所有这些抽象层级。为了解决这种偏差，研究组首先训练一个教师模型来模仿人类的判断，再通过微调，将其表征中与人类对齐的结构迁移至最先进的预训练视觉基础模型，从而优化其表征。

这些人类对齐模型在广泛的相似性任务中更准确地逼近人类行为和不确定性，包括跨越多层级语义抽象的人类判断数据集。它们在各种机器学习任务上也表现得更好，提高了泛化和分布外鲁棒性。

因此，向神经网络注入额外人类知识会产生一个两全其美的表征，既更符合人类的认知判断，又更实用，这为更强大、可解释和与人类一致的人工智能系统铺平了道路。

▲ Abstract： Deep neural networks have achieved success across a wide range of applications, including as models of human behaviour and neural representations in vision tasks. However, neural network training and human learning differ in fundamental ways, and neural networks often fail to generalize as robustly as humans do, raising questions regarding the similarity of their underlying representations. We need to determine what is missing for modern learning systems to exhibit more human-aligned behaviour. Here we highlight a key misalignment between vision models and humans: whereas human conceptual knowledge is hierarchically organized from fine- to coarse-scale distinctions, model representations do not accurately capture all these levels of abstraction. To address this misalignment, we first train a teacher model to imitate human judgements, then transfer human-aligned structure from its representations to refine the representations of pretrained state-of-the-art vision foundation models via fine-tuning. These human-aligned models more accurately approximate human behaviour and uncertainty across a wide range of similarity tasks, including a dataset of human judgements spanning multiple levels of semantic abstractions. They also perform better on a diverse set of machine learning tasks, increasing generalization and out-of-distribution robustness. Thus, infusing neural networks with additional human knowledge yields a best-of-both-worlds representation that is both more consistent with human cognitive judgements and more practically useful, paving the way towards more robust, interpretable and human-aligned artificial intelligence systems.   

 材料科学Materials Science   

Atomically resolved edges and defects in lead halide perovskites 

卤化铅钙钛矿的原子分辨边缘和缺陷分析    

▲ 作者：Biao Yuan, Zeyu Wang, Shuchen Zhang, Christoph Hofer, Chuang Gao, Tamazouzt Chennit, et al.

▲链接：

 https://www.nature.com/articles/s41586-025-09693-6 

▲摘要：虽然边缘和缺陷只构成晶体材料的一小部分，但它们对材料的性能产生巨大的影响。有机—无机卤化物钙钛矿是颇具前景的下一代半导体材料，具有优越的成本效益和有趣的光电性能，但因其极端敏感性，获得边缘的清晰图像仍颇具挑战。

研究组采用真正高速超低剂量四维扫描透射电子显微镜技术，结合剂量分割法，实现了迄今所知最低剂量的原子分辨率叠层成像，不仅揭示了卤化物钙钛矿边缘的精细原子结构，还揭示了边缘结构动力学。

研究组在甲基铵铅碘（MAPbI₃）中观察到大多数甲基铵（MA）和碘（I）在边缘的末端结构，并发现其边缘和内部缺陷的损坏率取决于存在空位的浓度和类型，特别是碘空位的优势与更高的损坏率相关。 

▲ Abstract： Although edges and defects constitute only a small fraction of crystalline materials, they exert an outsized impact on a material′s properties. Organic–inorganic halide perovskites are promising next-generation semiconductor materials with superior cost effectiveness and interesting optoelectronic properties. However, clear images of their edges have remained challenging to obtain owing to their extreme sensitivity. Using truly high-speed ultralow-dose four-dimensional scanning transmission electron microscopy with dose fractionation, we perform ptychography at, to our knowledge, the lowest-dose atomic resolution to date, revealing not only the detailed atomic structure of the edges of a halide perovskite but also their structural dynamics. A majority methylammonium (MA) and iodine (I) edge termination is observed in methylammonium lead iodide (MAPbI₃), and the damage rate of its edges and internal defects is found to depend on the concentration and type of vacancies present, with a preponderance of I vacancies in particular correlating with higher rates of damage.    

Silicon solar cells with hybrid back contacts 

杂化背接触结构硅太阳能电池    

▲ 作者：Genshun Wang, Mingzhe Yu, Hua Wu, Yunpeng Li, Lei Xie, Junzhe Wei, et al.

▲链接：

https://www.nature.com/articles/s41586-025-09681-w 

▲摘要：硅太阳能电池对可持续能源至关重要，但仍受到效率损失的限制，特别是在填充因子方面。

研究组开发了一种杂化背接触结构太阳能电池，其结合了先进的全表面钝化和激光加工隧穿接触。该策略实现了27.81%的功率转换效率，接近理论极限的95%。通过整合高低温工艺，研究组抑制了复合，提高了接触性能，实现了87.55%（接近理论极限98%）的填充因子。

该模型将理想因子与载流子损失机制联系起来，阐明了在体相和表面的载流子复合，以及由复合引起的关键填充因子损失。这些创新为可扩展、高效的硅光伏电池提供了实验和理论上的进展。

▲ Abstract： Silicon solar cells are essential for sustainable energy but remain limited by efficiency losses, particularly in the fill factor. Here we develop a hybrid interdigitated back-contact solar cell that combines advanced all-surface passivation with laser-treated tunnelling contacts. This approach achieves a power conversion efficiency of 27.81%, approaching 95% of the theoretical limit. By integrating high- and low-temperature processes, we suppress recombination and enhance contact performance, achieving a fill factor of 87.55%—nearly 98% of the theoretical limit. A model links the ideality factor to carrier loss mechanisms, elucidating carrier recombination in both the bulk and the surface and clarifies key fill factor losses owing to recombination. These innovations provide both experimental and theoretical advances towards scalable, high-efficiency silicon photovoltaics.    

化学Chemistry   

A molecularly impermeable polymer from two-dimensional polyaramids 

二维聚芳酰胺制备分子不渗透性聚合物    

▲ 作者：Cody L. Ritt, Michelle Quien, Zitang Wei, Hagen Gress, Mohan T. Dronadula, Kaan Altmisdort, et al.

▲链接：

https://www.nature.com/articles/s41586-025-09674-9 

▲摘要：所有聚合物都因缠结聚合物链的自由体积而表现出气体渗透性。相比之下，包括石墨烯在内的二维（2D）材料密堆可表现出分子不渗透性。溶液合成的2D聚合物通过缩聚反应表现出分子不渗透性是一个长期目标。

研究组展示了自支撑、自旋涂覆的2D聚芳酰胺纳米膜，其氮气渗透率低于3.1×10^-9 Barrer，比每种已知聚合物低近4个数量级，对其他测试气体（氦气、氩气、氧气、甲烷和六氟化硫）亦是如此。在纳米膜涂层微孔的加压过程中，光学干涉技术可实现对机械敏感边缘开合状态的测量，并形成稳定期超过三年的鼓泡结构。 

这一发现使2D聚合物谐振器具有高谐振频率（约8 MHz）和高达537的品质因数，类似于石墨烯。60 nm气敏钙钛矿涂层使晶格退化率降低了14倍，氧气渗透率为3.3×10^-8 Barrer。分子不渗透性聚合物有望实现下一代屏障，具有合成可加工性、化学适应性，并以最少的材料来最大化分子排斥率，最终实现可持续发展目标。 

▲ Abstract： All polymers exhibit gas permeability through the free volume of entangled polymer chains. By contrast, two-dimensional (2D) materials including graphene stack densely and can exhibit molecular impermeability. Solution-synthesized 2D polymers that exhibit the latter by poly-condensation have been a longstanding goal. Herein, we demonstrate self-supporting, spin-coated 2D polyaramid nanofilms that exhibit nitrogen permeability below 3.1 × 10^-9Barrer, nearly four orders of magnitude lower than every class of existing polymer, and similar for other gases tested (helium, argon, oxygen, methane and sulfur hexafluoride). Optical interference during the pressurization of nanofilm-coated microwells allows measurement of mechanosensitive rim opening and sealing, creating gas-filled bulges that are stable exceeding three years. This discovery enables 2D polymer resonators with high resonance frequencies (about 8 MHz) and quality factors up to 537, similar to graphene. A 60-nm coating of air-sensitive perovskites reduces the lattice degradation rate 14-fold with an oxygen permeability of 3.3 × 10^-8 Barrer. Molecularly impermeable polymers promise the next generation of barriers that are synthetically processable, chemically amenable and maximize molecular rejection with minimal material, ultimately advancing sustainability goals.    

A multimodal robotic platform for multi-element electrocatalyst discovery 

多模态机器人平台助力多元素电催化剂发现    

▲ 作者：Zhen Zhang, Zhichu Ren, Chia-Wei Hsu, Weibin Chen, Zhang-Wei Hong, Chi-Feng Lee, et al.

▲链接：

 https://www.nature.com/articles/s41586-025-09640-5 

▲摘要：“科学人工智能”的目标之一是通过现实世界的实验发现定制材料。人们已在计算预测和材料合成自动化方面取得了开创性进展。然而，大多数材料实验仍局限于使用单模态主动学习方法，依赖于单一数据流。人工智能解释实验复杂性的潜力在很大程度上尚未得到开发。

研究组介绍了真实世界实验科学家助驾（CRESt），这是一个将大型多模态模型（包括化学成分、文本嵌入和微观结构图像）与知识辅助贝叶斯优化和机器人自动化集成为一体的平台。CRESt采用基于知识嵌入的搜索空间缩减和自适应探索开发策略来加速材料设计、高通量合成和表征以及电化学性能优化。 

CRESt可通过摄像头进行监控，并生成视觉语言模型驱动的假设，以诊断和纠正实验异常。CRESt应用于甲酸盐电化学氧化，在3个月内分析了900多种催化剂化学性质，进行了3500次电化学测试，最终在八元化学空间（Pd-Pt-Cu-Au-Ir-Ce-Nb-Cr）中确定了一种最先进的催化剂，其成本效益提高了9.3倍。 

▲ Abstract： One of the goals of ‘AI for Science’ is to discover customized materials through real-world experiments. Pioneering advances have been made in computational predictions and the automation of materials synthesis. Yet most materials experimentation remains constrained to using unimodal active learning approaches, relying on a single data stream. The potential of artificial intelligence to interpret experimental complexity remains largely untapped. Here we present Copilot for Real-world Experimental Scientists (CRESt), a platform that integrates large multimodal models (incorporating chemical compositions, text embeddings and microstructural images) with knowledge-assisted Bayesian optimization and robotic automation. CRESt uses knowledge-embedding-based search space reduction and adaptive exploration–exploitation strategy to accelerate materials design, high-throughput synthesis and characterization, and electrochemical performance optimization. CRESt enables monitoring with cameras and the generation of vision-language-model-driven hypotheses to diagnose and correct experimental anomalies. Applied to electrochemical formate oxidation, CRESt explored more than 900 catalyst chemistries and 3,500 electrochemical tests within 3 months, identifying a state-of-the-art catalyst in the octonary chemical space (Pd–Pt–Cu–Au–Ir–Ce–Nb–Cr) that exhibits a 9.3-fold improvement in cost-specific performance.